The evolution of entity linking tasks in knowledge graphs from a single modality to a multi-modal framework has addressed the limitations associated with the singularity of data in single-modal approaches and mitigated the issue of compromising the inherent unique spatial characteristics of each modality due to the heterogeneity of embedding spaces in multi-modal integrations. However, despite these advancements, the efficiency of algorithms employed in the latest methods remains suboptimal, preventing the achievement of the most desirable outcomes. This paper addresses the challenge of improving multi-modal knowledge graph link prediction by proposing a novel method, OTGNET. Existing approaches struggle with inefficient integration across modalities, leading to suboptimal performance. Our approach uses neural networks to dynamically configure the optimal transport matrix, optimized with the Generalized Sliced Wasserstein Distance (GSWD). This preserves the unique structure of each modality while enhancing alignment and representation across modalities. Evaluations on the WN9-IMG and FB-IMG benchmarks show that OTGNET achieves a 15% improvement in accuracy over previous state-of-the-art methods, demonstrating the effectiveness of our approach.